A New Approach in Determining Lateral Facial Attractiveness

A New Approach in Determining Lateral Facial Attractiveness Koohyar Karimi a, Zlatko Devcic a, David Avila a, Natalie Popenko a, Brian Wong abc Beckman Laser Institute, 1002 Health Sciences Road, Irvine, CA 92612 b Department of Otolaryngology- Head and Neck Surgery, University of California Irvine, 101 The City Drive Bld. 25 Rt. 81, Orange, CA 92868, c Department of Biomedical Engineering Rockwell Engineering Center 204, University of California Irvine, Irvine CA 92612

1. Abstract

7. Synthetic Lateral Images were successfully created

4. Synthetic Lateral Facial Analysis

Introduction: The current literature on facial attractiveness focuses on anterior-posterior facial portraits, with lateral facial analysis limited to comparing facial attractiveness scores with various facial measurements. Here we use a novel approach to more rigorously study lateral facial attractiveness by combining morphing software and a genetic algorithm with web-based facial attractiveness scoring to evolve attractive lateral facial images.

An example of synthetic F3 morph created from two synthetic F2 morphed images. Each synthetic image is a composite (50/50) of two images.

7 6 5 Scores

Objective: The objectives of this study were to: 1) identify the key lateral facial landmarks that produce realistic lateral facial images; and 2) determine if a genetic algorithm combined with morphing software can progressively evolve lateral facial attractiveness.

Average Beauty Scores

Image 1

Methods: A cohort of lateral facial portraits were selectively paired by a genetic algorithm biased towards more attractive faces, and “bred” with morphing software to create a cohort of faces more attractive than the original. By repeating this process facial attractiveness was “evolved” through several cohorts.

Synthetic Image

Image 2

Attractive

4

Unattractive

3 2 1

User Defined Registry Points

0 1

2

3

4

5

Generations

Results: Key facial landmarks are: trichion to glabella, nasion to tip of nose, subnasale to labrale inferius, and pogonion to menton. Facial attractiveness scores increased in each successive cohort.

Attractiveness Score Calculation Conclusion: Using these landmarks and methodologies, realistic lateral facial portraits were created and progressively increased in facial attractiveness. This technique is a robust alternative to traditional approaches in the analysis of lateral facial attractiveness.

2.

Each image was posted on an internet-based rating website (Hotornot.com) until each face was rated 200-300 times. The website provides scores for each image between 1 and 10.

Background

Beauty is an elusive quality to measure, and attempts at quantifying it have in general failed due to the intrinsic subjective nature of this quality. During the Renaissance, da Vinci and Durer analyzed the frontal view of the face and developed the Classical Cannons which have permeated art, science, and medicine for the past five centuries. Modern approaches to study facial beauty have used focus group evaluations to rate the attractiveness of facial photographs and then correlate these attractiveness scores with various linear or angular measurements of the face. Focus groups have included trained experts and lay groups of evaluators that score faces based on a numerical scale. These studies are not only labor and time intensive, but also limit the size of subject pool and evaluators as they must be performed by 1) using focus groups to rate and evaluate photographs and 2) obtaining study participants willing to allow their facial features to be meticulously measured and then scrutinized. By utilizing morphing technology to create facial photographs with web-based large scale population facial scoring techniques, our study can provide a more rigorous quantitative and qualitative approach to define and evolve facial attractiveness.

3. Mechanical Design Genetic Breeding Algorithm

9.1 6.2

6.2

There are different approaches to create synthetic lateral images. Using a genetic breeding algorithm (see above) biased towards more attractive facial beauty scores, can selectively allow for cohort faces to be more attractive. Likewise, a selective breeding algorithm can randomly pair attractive and unattractive faces together to “breed” a cohort of images that are more attractive. (see below)

9.3

Lateral facial images were placed on a rating website (hotornot.com) to obtain facial attractiveness scores

9.3

P

P

F1

F1

F2

F2

F3

F3

F4

F4

Selection Pressure

6. Results 62 synthetic lateral images were created from digital photographs of volunteers by morphing software. The 16 highest scoring images (the “most attractive” top 50th percentile) of the parent generation were randomly morphed together in order to produce the F1 generation of 8 “attractive” offspring synthetic lateral images. Images in the bottom 50th percentile were also morphed to produce the F1 generation of 8 “unattractive” offspring. This process was repeated for 3 more generations in order to compile each synthetic image in an attractive or unattractive category.

Selective Breeding Algorithm

1- trichion to glabella 2- nasion to subnasale 3- subnasale to vermillion border 4- stomion to menton 5- pogonion to the menton 6- outline of lip 7- exterior ear boundary

“breeds”

Gene Pool

Highest Scored Attractive Lateral Images

Reference points used to create a realistic synthetic lateral image

3.8

3.8

“fails”

Lowest Scored Unattractive Lateral Images

Key features of eye area: 1- outline of eyebrow 2- outline of eyelashes 3- lateral palpebral commissure 4- upper eyelid outline 5- lower eyelid outline 6- pupil outline 7- lateral scelra

Random Number Generator

Random Selection

9.1

5. Registry point assignment to distinct facial features

Scores of the lateral images were used to categorize the synthetic images in their “attractive” or “unattractive” categories.

Aim 1: Critical registry points were successfully defined to create a synthetic lateral image. (Figure 5) Aim 2: Synthetic images created in all generations of either attractive or unattractive categories maintained clear defined contours of all facial features. (Figure 7)

8. Conclusion The methodology to identify the critical registry points required to create a synthetic image was successful in all generations of either attractive or unattractive categories. This process has also identified crucial distinct facial features required to create a realistic synthetic image in each generation.

9. Future Work Is it possible to create an automatic morphing technology? Expand pilot study to quantify distinct lateral facial features. Expand on different ethnic-orientated lateral images.